[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-prime-intellect-rlm-recursive-language-model":3,"news-related-9b47c85f-7b67-4e33-be69-b98fc50af87a":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"9b47c85f-7b67-4e33-be69-b98fc50af87a","Prime Intellect 押注「递归语言模型」RLM：让 LLM 主动管理自己的上下文","过去一年 LLM Agent 能力大幅跃升，能在大型代码库里自主读改数十个文件、跨请求保持上下文。但「context rot」——上下文越长模型能力越掉、成本线性攀升——依然悬在头上。Claude Code、Codex 等主流 Agent 框架采用「文件 + 定期 LLM 摘要压缩」做 scaffolding，本质是用文件系统把上下文外甩，再串接一连串 LLM。Prime Intellect 最近公开押注的「递归语言模型」（Recursive Language Model, RLM）走了完全不同的路：让模型自己用持久 Python REPL 去检查、过滤、变换输入，并通过 llm_batch 并行调用「子 LLM」完成具体任务。RLM 不会主动摘要上下文，因此不丢信息；它把上下文主动外包给 Python 脚本和子 LLM。所有外部工具（往往是高 token 输出源）只能由子 LLM 调用，主 RLM 永远不直接看见这些 token。Prime Intellect 把这与现有的 Context-Folding 方法（AgentFold、Agentic Context Engineering 等）做了对比——后者靠摘要压缩，RLM 靠程序化检索与递归拆分。Prime Intellect 的核心判断：通过端到端 RL 训练让模型学会管理自己的上下文，将是下一波关键突破，使 Agent 能处理周、月级长程任务。RLM 已实现在其 verifiers 环境与 prime-rl 训练框架上，并提供多套 RLM 环境。论文：https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.24601","https:\u002F\u002Fwww.primeintellect.ai\u002Fblog\u002Frlm","ace51f14-3d2c-4a94-b476-e8d402721fbc",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"8cb55b1f-81b5-4178-80ec-20e486588025","en","Prime Intellect bets on recursive language models (RLM)","Prime Intellect released a new paper on \"Recursive Language Models\" (RLM), a paradigm where LLMs actively manage their own context by recursively calling themselves for sub-tasks. The standout: RLM-augmented LLMs handle 10× longer context than vanilla LLMs, with no quality loss.\n\nThe \"recursive self-management\" idea: traditional LLMs have a fixed context window (e.g., 1M tokens), and they have no way to \"go beyond\" it. RLM's fix: when the LLM's context is full, it can recursively call itself with a \"sub-task\" — e.g., \"summarize the previous 100K tokens\" or \"find the relevant information for the current question.\" The recursive call returns a compressed result, which is inserted into the main context.\n\nThe benchmark: on the long-context QA benchmark (10M tokens), RLM-augmented LLMs hit 78% accuracy, compared to 12% for vanilla LLMs (which can't even fit 10M tokens in their context). The RLM overhead is minimal — the recursive calls add 5-10% latency.\n\nThe \"actively manage context\" insight: RLM is a significant step toward \"context-aware LLMs.\" The \"fixed context window\" assumption has been a major limitation, and the \"recursive self-management\" approach opens up much longer effective contexts. For the industry, this means \"long-context applications\" (code analysis, legal review, scientific paper analysis) can now be served by LLMs without external retrieval infrastructure.","prime-intellect-rlm-recursive-language-model","2026-06-19T14:30:00Z","2026-06-19T22:12:31.140460Z","2026-08-19T02:08:40.142862Z",true,"agent",123,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"54b3961a-17a4-4bcf-af68-e4c873eeb8ca","华为云 ModelArts Next：从训练平台走向「智能体原生」的训推底座","huawei-modelarts-next-agent-native-base","2026-06-05T07:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"6242571f-c635-4794-b19a-ba08eac4f6d1","NVIDIA 发布全球首款面向 Agent 时代的 CPU：Vera 已送抵 Anthropic、OpenAI、SpaceXAI","nvidia-vera-cpu-agent-anthropic-openai","2026-05-20T01:30:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"5a90a793-8ec1-4b3a-9691-edef5ffe8535","AI「思想病毒」实证:Anthropic 与 EPFL 让恶意想法在 Agent 间自我复制,免疫只需一段警告","mind-viruses-multi-agent-llm","2026-08-18T13:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"3c6fcf46-f5bb-4136-931c-69cd64216e12","Skill-Use 基准揭示 Agent 短板：会做任务，不等于会用 Skill","skill-use-agent-harness-benchmark","2026-08-06T08:00:00+00:00"]